Summary
This paper (introducing ERVLA) revisits embodied chain-of-thought (CoT) reasoning for VLA models at large scale, building the largest embodied CoT corpus to date — 978,743 trajectories, 226.3M samples, 2,592.5 hours of robot data — to systematically study what kind of reasoning actually helps manipulation and how it should be used at inference time. Its central finding is that grounded, concrete reasoning (end-effector movement descriptions, image-space trajectories) helps substantially, high-level semantic reasoning alone barely helps, and naively decoding CoT as an autoregressive action prefix hurts scalability due to compounding inference errors.
Key Contributions
- The largest embodied CoT dataset assembled to date, enabling much more statistically reliable conclusions about what kinds of reasoning traces transfer to policy improvement than prior smaller-scale CoT-for-robotics studies.
- An empirical finding that concrete, action-grounded reasoning (motion descriptions, image-space trajectory sketches) drives most of the benefit, while abstract high-level semantic reasoning contributes only marginal gains — a useful corrective to assumptions that “more reasoning is better.”
- Identifies that using CoT as an explicit autoregressive prefix at inference time doesn’t scale reliably because errors in the generated reasoning compound into the eventual action, degrading rather than improving performance at scale.
- Introduces reasoning-dropout training: the model absorbs rich reasoning traces during training but predicts actions directly at inference without decoding CoT, getting the training-time benefit of reasoning supervision without paying an inference-time compounding-error cost.
- Achieves strong results: 86.9% success on LIBERO-Plus and 53.2% on VLABench, indicating good out-of-distribution generalization relative to other reported baselines on these benchmarks.
Strengths
- The scale of the CoT corpus (nearly 1M trajectories, 2.6K hours) gives this study much more empirical weight than typical ablations on a few thousand trajectories, making its negative results (high-level reasoning doesn’t help; explicit CoT decoding doesn’t scale) more credible.
- The reasoning-dropout technique is a clever, practically important resolution to the tension between “reasoning helps during training” and “explicit reasoning decoding hurts at inference,” and is likely reusable by other embodied CoT approaches.
- Strong, clearly reported benchmark numbers (LIBERO-Plus, VLABench) situate the contribution concretely relative to the field’s standard generalization tests.
Weaknesses
- The negative finding about high-level semantic reasoning may be specific to how “high-level reasoning” was operationalized/labeled in their CoT corpus construction — a differently curated high-level reasoning signal might behave differently, so the generality of this conclusion is somewhat dependent on their labeling scheme.
- Because reasoning is dropped at inference, the model loses the interpretability benefit that was one of the original selling points of chain-of-thought for robotics (being able to inspect why the policy is doing what it’s doing at deployment time).
- Building a corpus of this scale (978K trajectories with reasoning annotations) likely required either automated annotation pipelines (e.g., via a large VLM labeling trajectories) whose own error/bias could propagate into the “ground-truth” reasoning supervision — the annotation quality-control process isn’t detailed here.
Open Questions
- Would the “concrete grounding beats abstract reasoning” finding hold for more genuinely novel or compositional tasks that plausibly require the kind of higher-level planning abstract reasoning is meant to provide?
- Is there some intermediate approach (partial/compressed reasoning decoding, or reasoning used only for uncertain/hard sub-steps) that could recover some interpretability benefit without the full compounding-error cost?
- How does the reasoning-dropout model’s performance change with corpus size — is the ~1M-trajectory scale actually necessary, or would a much smaller curated CoT corpus achieve similar gains?
Significance
By combining an unusually large embodied CoT corpus with a careful ablation of what reasoning content and inference-time usage strategy actually helps, this paper delivers some of the most rigorous evidence yet on how (and whether) chain-of-thought reasoning should be incorporated into VLA training and deployment — directly informing the design of future reasoning-augmented manipulation policies.